Magnetic resonance imaging radiomics-based machine learning model for preoperative prediction of pathological grade in pancreatic cancer: A preliminary study.

[BACKGROUND] To explore the value of MRI radiomics-based machine learning models for predicting the pathological grade of pancreatic cancer preoperatively.

[METHODS] 125 patients with pathologically confirmed pancreatic cancer who underwent preoperative MRI were retrospectively enrolled. The primary cohort was randomized in an 8:2 ratio into a training cohort (n = 100) and a validation cohort (n = 25). 1316 radiomics features were extracted from contrast-enhanced TWI arterial phase (AP) or portal venous phase (PVP) images, respectively. After feature reduction and filtering, the best features were selected to construct machine learning models (K-nearest neighbor, KNN; support vector machine, SVM; logistic regression, LR; random forest, RF). Finally, the performance of these models was evaluated using the receiver operating characteristic curve (ROC).

[RESULTS] There were no statistical differences in clinical characteristics between the low-grade and high-grade cohorts ( > 0.05). The best radiomics features selected from the AP, PVP and AP+PVP images were 6, 6 and 10, respectively. Among the four models, the LR machine learning model achieved the best predictive performance. The distribution of the Radscore values was clinically significant between the low-grade and high-grade groups both in the training cohort (median, 0.26 vs 0.99;  < 0.001) and validation cohort (median, 0.63 vs 1.48;  = 0.011). LR model of AP+PVP performed the best with AUC value of 0.81 (95 % CI: 0.72-0.91) for the training cohort and 0.82 (95 % CI: 0.62-1.00) for the validation cohort.

[CONCLUSIONS] MRI radiomics-based machine learning model is a potential non-invasive method to predict the pathological grade of pancreatic cancer.